Transcript References

Computational Drug Design
Noel M. O’Boyle
Apr 2010
Postgrad course on Comp Chem
References
• Computational Drug Design, David C.
Young
• My EBI lecture on protein-ligand
docking, http://slideshare.net/baoilleach
• An Introduction to Cheminformatics, AR
Leach, VJ Gillet
Rational Drug Design
• Use knowledge of protein or ligand structures
– Does not rely on trial-and-error or screening
– Computer-aided drug design (CADD) now plays an
important role in rational design
• Structure-based drug design
– Uses protein structure directly
– CADD: Protein-ligand docking
• Ligand-based drug design
– Derive information from ligand structures
– Protein structure not always available
• 40% of all prescription pharmaceuticals target GPCRs
– Protein structure has large degree of flexibility
• Structure deforms to accommodate ligands or gross
movements occur on binding
– CADD: Pharmacophore approach, Quantitative structureactivity relationship (QSAR)
Unknown
protein structure
Known protein
structure
Computer-aided drug design (CADD)
Known ligand(s)
No known ligand
Structure-based drug
design (SBDD)
De novo design
Protein-ligand docking
Ligand-based drug design
(LBDD)
1 or more ligands
• Similarity searching
Several ligands
• Pharmacophore searching
Many ligands (20+)
• Quantitative Structure-Activity
Relationships (QSAR)
CADD of no use
Need experimental
data of some sort
Can apply ADMET
filters
Virtual screening
• Virtual screening is the computational or in silico
analogue of biological screening
• The aim is to score, rank or filter a set of structures
using one or more computational procedures
• It can be used
–
–
–
–
to help decide which compounds to screen (experimentally)
which libraries to synthesise
which compounds to purchase from an external company
to analyse the results of an experiment, such as a HTS run
Virtual screening
AR Leach, VJ Gillet, An Introduction to Cheminformatics
What is a Pharmacophore?
• Two somewhat distinct usages:
•
•
That substructure of a molecule that is responsible for its pharmacological
activity (c.f. chromophore)
A set of geometrical constraints between specific functional groups that
enable the molecule to have biological activity
Bojarski, Curr. Top. Med. Chem. 2006, 6, 2005.
Overview of Pharmacophore-based Drug Design
Activity data
Generate
pharmacophore
Test activity
Search compound
library for actives
Buy or synthesise ‘hits’
See also John Van Drie’s http://pharmacophore.org
Pharmacophore generation and searching
• Protein structure not required
– There are also approaches that create pharmacophores
from the active site
• Assumes that all (or the majority) of the known actives
bind to the same location
• Pharmacophore generation
– Identify pharmacophoric features (hydrogen bond donors
and acceptors, lipophilic groups, charges)
– Find a geometrical arrangement of pharmacophoric features
that all actives that match with a low-energy conformation
• Pharmacophore searching
– Given a pharmacophore, find all molecules in a database
that can match it in a low-energy conformation
• Some pharmacophore software gives an estimate of activity, but
most just give true or false for a match
– Scaffold-hopping possible
• Doesn’t require structural similarity
• Just needs to match the pharmacophore
Protein-ligand docking
• A Structure-Based Drug Design (SBDD) method
– “structure” means “using protein structure”
• Computational method that mimics the binding of a ligand to a
protein
• Given...
• Predicts...
• The pose of the molecule in
the binding site
• The binding affinity or a
score representing the
strength of binding
Protein-ligand docking II
•
•
Typically, protein-ligand docking software consist of two main components
which work together:
1. Search algorithm
–
•
2. Scoring function
–
•
Calculates a score or binding affinity for a particular pose
The difficulty with protein–ligand docking is in part due to the fact that it
involves many degrees of freedom
–
–
–
•
Generates a large number of poses of a molecule in the binding site
The translation and rotation of one molecule relative to another involves six degrees of
freedom
There are in addition the conformational degrees of freedom of both the ligand and the
protein
The solvent may also play a significant role in determining the protein–ligand geometry
The search algorithm generates poses, orientations of particular
conformations of the molecule in the binding site
–
–
Tries to cover the search space, if not exhaustively, then as extensively as possible
There is a tradeoff between time and search space coverage
Ligand conformations
• Conformations are different three-dimensional structures of
molecules that result from rotation about single bonds
– That is, they have the same bond lengths and angles but different torsion
angles
• For a molecule with N rotatable bonds, if each torsion angle is
rotated in increments of θ degrees, number of conformations is
(360º/ θ)N
• Question
– If the torsion angles are incremented in steps of 30º, how many
conformations does a molecule with 5 rotatable bonds have,
compared to one with 4 rotatable bonds?
• Having too many rotatable bonds results in “combinatorial
explosion”
• Also ring conformations
Taxol
Types of search algorithms
• Classified based on the degrees of freedom that they
consider
• Rigid docking
– The ligand is treated as a rigid structure during the docking
• Only the translational and rotational degrees of freedom are considered
– To deal with the problem of ligand conformations, a large
number of conformations of each ligand are generated in
advance and each is docked separately
• Flexible docking is more common today
– Conformations of each molecule are generated on-the-fly by
the search algorithm during the docking process
– Avoids considering conformations that do not fit
– Exhaustive (systematic) searching computationally too
expensive as the search space is very large
– One common approach is to use stochastic search methods
• These don’t guarantee optimum solution, but good solution within
reasonable length of time
• Stochastic means that they incorporate a degree of randomness
• Includes genetic algorithms (GOLD), simulated annealing (AutoDock)
Handling protein conformations
• Most docking software treats the protein as rigid
– Rigid Receptor Approximation
• This approximation may be invalid for a particular
protein-ligand complex as...
– the protein may deform slightly to accommodate different
ligands (ligand-induced fit)
– protein side chains in the active site may adopt different
conformations
• Some docking programs allow
protein side-chain flexibility
– For example, selected side chains are
allowed to undergo torsional rotation
around acyclic bonds
– Increases the search space
• Larger protein movements can
only be handled by separate
dockings to different protein
conformations
The perfect scoring function will…
• Accurately calculate the binding affinity
– Will allow actives to be identified in a virtual screen
– Be able to rank actives in terms of affinity
• Score the poses of an active higher than poses of an
inactive
– Will rank actives higher than inactives in a virtual screen
• Score the correct pose of the active higher than an
incorrect pose of the active
– Will allow the correct pose of the active to be identified
• Broadly speaking, scoring functions can be divided into
the following classes:
– Forcefield-based
• Based on terms from molecular mechanics forcefields
• GoldScore, DOCK, AutoDock
– Empirical
• Parameterised against experimental binding affinities
• ChemScore, PLP, Glide SP/XP
– Knowledge-based potentials
• Based on statistical analysis of observed pairwise distributions
• PMF, DrugScore, ASP
Böhm’s empirical scoring function
•
In general, scoring functions assume that the free energy of binding can be
written as a linear sum of terms to reflect the various contributions to binding
•
•
Bohm, J. Comput.-Aided Mol. Des., 1994, 8, 243
Bohm’s scoring function included contributions
from hydrogen bonding, ionic interactions, lipophilic
interactions and the loss of internal conformational
freedom of the ligand.
•
The hydrogen bonding and ionic terms are both dependent on the geometry
of the interaction, with large deviations from ideal geometries (ideal distance R,
ideal angle α) being penalised.
The lipophilic term is proportional to the contact surface area (Alipo) between
protein and ligand involving non-polar atoms.
The conformational entropy term is the penalty associated with freezing
internal rotations of the ligand. It is largely entropic in nature. Here the value is
directly proportional to the number of rotatable bonds in the ligand (NROT).
The ∆G values on the right of the equation are all constants determined using
multiple linear regression on experimental binding data for 45 protein–ligand
complexes
•
•
•
–
Hence “empirical”
Pose prediction accuracy
• Given a set of actives with known crystal poses, can
they be docked accurately?
• Accuracy measured by RMSD (root mean squared
deviation) compared to known crystal structures
– RMSD = square root of the average of (the difference
between a particular coordinate in the crystal and that
coordinate in the pose)2
– Within 2.0Å RMSD considered cut-off for accuracy
– More sophisticated measures have been proposed, but are
not widely adopted
• In general, the best docking software predicts the
correct pose about 70% of the time
• Note: it’s always easier to find the correct pose when
docking back into the active’s own crystal structure
– More difficult to cross-dock
Assess performance of a virtual screen
• Need a dataset of Nact known actives, and inactives
• Dock all molecules, and rank each by score
• Ideally, all actives would be at the top of the list
– In practice, we are interested in any improvement over what
is expected by chance
• Define enrichment, E, as the number of actives found
(Nfound) in the top X% of scores (typically 1% or 5%),
compared to how many expected by chance
– E = Nfound / (Nact * X/100)
– E > 1 implies “positive enrichment”, better than random
– E < 1 implies “negative enrichment”, worse than random
• Why use a cut-off instead of looking at the mean rank
of the actives?
– Typically, the researchers might test only have the resources
to experimentally test the top 1% or 5% of compounds
• More sophisticated approaches have been developed
(e.g. BEDROC) but enrichment is still widely used
Preparing the protein structure
• The Protein Data Bank (PDB) is a repository of protein
crystal structures, often in complexes with inhibitors
• PDB structures often contain water molecules
– In general, all water molecules are removed except where it is
known that they play an important role in coordinating to the ligand
• PDB structures are missing all hydrogen atoms
– Many docking programs require the protein to have explicit
hydrogens. In general these can be added unambiguously, except
in the case of acidic/basic side chains
NH
N
• An incorrect assignment of protonation
states in the active site will give poor
NH
HN
results
+
R
• Glutamate, Aspartate have COO- or
COOH
R
N
H
N
– OH is hydrogen bond donor, O- is not
• Histidine is a base and its neutral form
has two tautomers
R
Preparing the protein structure
• For particular protein side chains, the PDB structure can
be incorrect
• Crystallography gives electron density, not molecular
structure
– In poorly resolved crystal structures of proteins, isoelectronic
groups can give make it difficult to deduce the correct structure
O
R
NH2
R
NH 2
N
R
R
O
N
N
N
• Affects asparagine, glutamine, histidine
• Important? Affects hydrogen bonding pattern
• May need to flip amide or imidazole
– How to decide? Look at hydrogen bonding pattern in crystal
structures containing ligands